Unmix song with extreme settings.... slow Zzzzzz

@Virtuoso Unfortunately Apple has dropped support for external GPU accelerator since they have their Apple Silicon machine. However I’m working on using some Metal acceleration with Apple, but it takes time for them to properly offer support for that…

1 Like

In the settings (in System tab) you can select the GPU for AI functions!

Only on PCs though - that option does not appear on Intel or Apple Silicon Macs.

To be clear, are you saying that the lack of Metal support in Spectralayers is down to Apple and that something has to change at their end before you are able to implement it?

@Virtuoso Yes. They first offered a technology called CoreML, which I tried to use for SpectraLayers, but it only worked for very simple models, mostly image-oriented models. Most audio models in SL didn’t work with that tech.
They also have a technology called MPS (Metal Performance Shader) which seem to handle more AI models… but the tools to convert AI models to this technology are pretty much broken. Which is a shame because the Apple Silicon hardware is pretty much AI-capable, but they seem to be mostly interested in Apple products using it, and not so much about other companies using it. The situation should improve over time though…

5 Likes

I tested the same file on Windows and Mac.
A finished mixed and mastered song with drums, vocals, guitars, bass, 2 saxophones, trumpet, trombone and horn, 4 min long.
Song unmix into all components with extreme setting.

Windows: i7-10700K 32GB DDR4 RTX3050
takes 10 min 30 sec!
For AI functions, the RTX in SL is selected! (as MegaPro described)

MacBook Air M1 base model: takes 42min 30 sec!

What’s noticeable on the Mac is that only the P-Cores run between 90-100%, the E-Cores between 5-30% and the GPU only runs at up to 10%!
Actually, the MacBook with its 16 neural engine cores should be much better suited for this type of task???
Something still needs to be adjusted in Spactralayers11…
I’m waiting for the first update!

@Cubase-Erklaerbaer see my reply just above your message

1 Like

Hey Robin,

any news on using Nvidia’s Tensor Cores with SL11?

Cheers,
Benji

@YourMusic.Pro See my message above, it’s now fully working :slight_smile:

The patch should be released next week !

6 Likes

Afaik, the current max amount of vram on a fanless gpu is 6gb (Palit kalmx 3050). Any suggestions for a fanned but equally silent one with more ram?

I am currently testing this very card. I’ll update with results as soon as the patch is released, and I’ve had time to test it. N.B. my expectations are modest, but the advantages (for me) are:

  • no fan
  • no special connectors
  • low power usage (<70W)

Robin,

I’m quite likely a little dense here, but I was referring to a statement in an earlier thread where you said that SL was only at this point using the regular CUDA cores and not the AI-specific Tensor cores. Has this changed?

I have the same (upgraded from kalmx 1650). No problems so far, but not a lot of gpu juice for extra processing. And of course, it makes zero noise.

1 Like

@YourMusic.Pro ah right, no it’s using regular CUDA cores. Using Tensor cores would require going through the CUDA API rather than the DirectML API, which prevent accelerating models on AMD or Intel GPUs.

That being said, CUDA cores already provide a tremendous speed boost over CPU. Tensor cores could potentially make it even faster, but that would also require special tweakings (converting the float32 models to float16 for instance). It might happen one day since ONNX Runtime (the library used to infer AI models in SpectraLayers) also support the CUDA API, but I prefer not to promess anything here :slight_smile:

4 Likes

You could consider MSI GeForce RTX 3060 GAMING X 12G

Also, some graphics cards stop their fans at low temps, or use adjustable fan curves. In my opinion, you are wasting a lot of potential with these lower - tier cards… My 4070 SUPER for example is dead silent in day to day operation, but will run circles around most 30XX cards…
I do realize that money is an issue, of course, but I use SL Pro often and work-related and am glad I did spend the money I eventually did… :wink: :+1:

Is fast, balanced, best from V10 the same as fast, high, extreme in V11, or is V10 best = V11 high and V11 extreme above V10 best?

SL10 Fast=SL11 Fast (except SL11 can also unmix Sax & Brass if needed)
SL10 Balanced<SL10 Best<SL11 High<SL11 Extreme

1 Like

Thanks for clarifying.

Btw. why is “Other” greyed out in SL11 and cannot be unticked?

In SL11 “Other” is the layer that will also receive everything that is unchecked in the dialog, so that it remains non-destructive (with the sum of the unmixed layers always resulting in the original).

So say you have a recording with vocals, piano and strings (strings usually goes to Other). If you uncheck “Piano”, then piano sounds will be added to “Other” along with the strings so that Vocals+Other=original layer.

1 Like